[mlir][sparse] allow all-dense annotated "sparse" tensor output

[aartbik]

Summary

This is a very careful start with alllowing sparse tensors at the
left-hand-side of tensor index expressions (viz. sparse output).
Note that there is a subtle difference between non-annotated tensors
(dense, remain n-dim, handled by classic bufferization) and all-dense
annotated "sparse" tensors (linearized to 1-dim without overhead
storage, bufferized by sparse compiler, backed by runtime support library).
This revision gently introduces some new IR to facilitate annotated outputs,
to be generalized to truly sparse tensors in the future.

[aartbik]

fixed old code

[aartbik]

variadic operands (forward looking to sparse generalization)

[aartbik]

removed code leftover

[aartbik]

rebase with main

[aartbik]

rebase

[gussmith23]

Some initial questions. The code looks good to me, though my code review muscles are out of shape, so I will take another pass after these comments!

mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td

125

Question for my own understanding: later, on line 248, you assert that there's just a single input argument (if I understand the code correctly). Will this always be the case? If so, why is this Variadic? Is ins always expected to be variadic?

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

285

Just to confirm: is this rewrite effectively a no-op? It just replaces the operator with the underlying pointer?

mlir/test/Dialect/SparseTensor/dense.mlir

118

Where would this fail, and why, exactly?

[aartbik]

mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td

125

Good question. This is forward looking to sparse outputs, where we will need to feed in all arrays that constitute the sparse tensor (see L 146 for an example of that). All this is pretty much moot with the current approach of lowering to runtime support library (where the operation is folded again into the opaque pointer) but alternative implementations are possible in the future (e.g. actual code generation for stuff that is currently backed by the runtime support library). In such cases, we must correctly reflect the fact that reconstructing a tensor from sparse storage schemes needs all arrays, even though currently it doesn't.

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

285

Yes, even the variadic (in a later revision) will just verify that they all map to the same pointer and lower it to the no-op.

However, if we abandon the runtime support backing with actual codegen, that may chane in the future, and it is better to correctly reflect the full depenence on the sparse storage scheme!

mlir/test/Dialect/SparseTensor/dense.mlir

118

L623 in Sparsification.cpp

Not having the same "buffer" going in and out would require an explicit memory allocation for the tensor into which we reconstruct the sparse tensor according to the op (see the two ways of inplace/not-in-place for non-annotated outputs).

[aartbik]

rebased with main

[gussmith23]

After these comments, I think I'm good! However as I said, someone else may want to review this, as I'm brand new to this code and frankly still do not understand everything that is happening.

mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td

125

Ah, I see! I didn't think about how you'd potentially need multiple memrefs to implement a sparse tensor.

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

381–385

750

Should this comment be changed? Seems like this supports sparse tensors as well now?

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

750

Good catch. Changed.

[aartbik]

changed comment, rebased with main

[bixia]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

278

s/fold/folds/

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

730–742

It seems to me that this code block calculating the args for load is repeated for store. If I am right here, we should find a way to put this to a subroutine for reuse.

[aartbik]

Thanks all!

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

730–742

Yes, with this move towards allowing sparse stores, the code is much more similar to before. I kept it like this to make the diffs smaller, but I will unify in next CL that touches this!

mlir/include/mlir/Dialect/SparseTensor/IR/SparseTensorOps.td

 
 def SparseTensor_ToPointersOp : SparseTensor_Op<"pointers", [NoSideEffect]>,
     Arguments<(ins AnyTensor:$tensor, Index:$dim)>,
     Results<(outs AnyStridedMemRefOfRank<1>:$result)> {
   let summary = "Extract pointers array at given dimension from a tensor";
   let description = [{
      Returns the pointers array of the sparse storage scheme at the
      given dimension for the given sparse tensor. This is similar to the
-     `buffer_cast` operation in the sense that it provides a bridge
+     `memref.buffer_cast` operation in the sense that it provides a bridge
      between a tensor world view and a bufferized world view. Unlike the
-     `buffer_cast` operation, however, this sparse operation actually
+     `memref.buffer_cast` operation, however, this sparse operation actually
      lowers into a call into a support library to obtain access to the
      pointers array.
 
      Example:
 
     ```mlir
     %1 = sparse_tensor.pointers %0, %c1
        : tensor<64x64xf64, #CSR> to memref<?xindex>
     ```
   }];
   let assemblyFormat = "$tensor `,` $dim attr-dict `:` type($tensor)"
       " `to` type($result)";
 }
 
 def SparseTensor_ToIndicesOp : SparseTensor_Op<"indices", [NoSideEffect]>,
     Arguments<(ins AnyTensor:$tensor, Index:$dim)>,
     Results<(outs AnyStridedMemRefOfRank<1>:$result)> {
   let summary = "Extract indices array at given dimension from a tensor";
   let description = [{
      Returns the indices array of the sparse storage scheme at the
      given dimension for the given sparse tensor. This is similar to the
-     `buffer_cast` operation in the sense that it provides a bridge
+     `memref.buffer_cast` operation in the sense that it provides a bridge
      between a tensor world view and a bufferized world view. Unlike the
-     `buffer_cast` operation, however, this sparse operation actually
+     `memref.buffer_cast` operation, however, this sparse operation actually
      lowers into a call into a support library to obtain access to the
      indices array.
 
      Example:
 
     ```mlir
     %1 = sparse_tensor.indices %0, %c1
        : tensor<64x64xf64, #CSR> to memref<?xindex>
     ```
   }];
   let assemblyFormat = "$tensor `,` $dim attr-dict `:` type($tensor)"
       " `to` type($result)";
 }
 
 def SparseTensor_ToValuesOp : SparseTensor_Op<"values", [NoSideEffect]>,
     Arguments<(ins AnyTensor:$tensor)>,
     Results<(outs AnyStridedMemRefOfRank<1>:$result)> {
   let summary = "Extract numerical values array from a tensor";
   let description = [{
      Returns the values array of the sparse storage scheme for the given
      sparse tensor, independent of the actual dimension. This is similar to
-     the `buffer_cast` operation in the sense that it provides a bridge
+     the `memref.buffer_cast` operation in the sense that it provides a bridge
      between a tensor world view and a bufferized world view. Unlike the
-     `buffer_cast` operation, however, this sparse operation actually
+     `memref.buffer_cast` operation, however, this sparse operation actually
      lowers into a call into a support library to obtain access to the
      values array.
 
      Example:
 
     ```mlir
     %1 = sparse_tensor.values %0 : tensor<64x64xf64, #CSR> to memref<?xf64>
     ```
   }];
   let assemblyFormat = "$tensor attr-dict `:` type($tensor) `to` type($result)";
 }
 
+def SparseTensor_ToTensorOp : SparseTensor_Op<"tensor", [NoSideEffect]>,
+    Arguments<(ins Variadic<AnyStridedMemRefOfRank<1>>:$memrefs)>,
+    Results<(outs AnyTensor:$result)> {
+  let summary = "Reconstructs tensor from arrays(s)";
+  let description = [{
+    Reconstructs the sparse tensor from the sparse storage scheme array(s).
+    This is similar to the `memref.load` operation in the sense that it
+    provides a bridge between a bufferized world view and a tensor world
+    view. Unlike the `memref.load` operation, however, this sparse operation
+    is used only temporarily to maintain a correctly typed intermediate
+    representation during progressive bufferization. Eventually the operation
+    is folded away.
+
+    The input arrays are defined unambigously by the sparsity annotations
+    (pointers and indices for overhead storage in every compressed dimension,
+    followed by one final values array).
+
+    Examples:
+
+    ```mlir
+    %1 = sparse_tensor.tensor %0 : memref<?xf64> to tensor<64x64xf64, #Dense>
+
+    %3 = sparse_tensor.tensor %0, %1, %2 :
+       memref<?xindex>, memref<?xindex>, memref<?xf32> to tensor<10x10xf32, #CSR>
+    ```
+  }];
+  let assemblyFormat = "$memrefs attr-dict `:` type($memrefs) `to` type($result)";
+}
+
 #endif // SPARSETENSOR_OPS

mlir/lib/Dialect/SparseTensor/IR/SparseTensorDialect.cpp

if (!getSparseTensorEncoding(op.tensor().getType()))
     return op.emitError("expected a sparse tensor to get values");
   RankedTensorType ttp = op.tensor().getType().cast<RankedTensorType>();
   MemRefType mtp = op.result().getType().cast<MemRefType>();
   if (ttp.getElementType() != mtp.getElementType())
     return op.emitError("unexpected mismatch in element types");
   return success();
 }
 
+// TODO: generalize this beyond all-dense linearized "sparse" tensors
+static LogicalResult verify(ToTensorOp op) {
+  if (op.getNumOperands() != 1)
+    return op.emitError("expected single values array");
+  if (auto e = getSparseTensorEncoding(op.result().getType())) {
+    auto dlt = e.getDimLevelType();
+    for (unsigned i = 0, sz = dlt.size(); i < sz; i++) {
+      if (dlt[i] != SparseTensorEncodingAttr::DimLevelType::Dense)
+        return op.emitError("unexpected non-dense dimension");
+    }
+    return success();
+  }
+  return op.emitError("expected a sparse tensor as result");
+}
+
 //===----------------------------------------------------------------------===//
 // TensorDialect Methods.
 //===----------------------------------------------------------------------===//
 
 void SparseTensorDialect::initialize() {
   addAttributes<
 #define GET_ATTRDEF_LIST
 #include "mlir/Dialect/SparseTensor/IR/SparseTensorAttrDefs.cpp.inc"

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

matchAndRewrite(ToValuesOp op, ArrayRef<Value> operands,
     else
       return failure();
     rewriter.replaceOpWithNewOp<CallOp>(
         op, resType, getFunc(op, name, resType, operands), operands);
     return success();
   }
 };
 
+/// Sparse conversion rule for tensor reconstruction.
+class SparseTensorToTensorConverter : public OpConversionPattern<ToTensorOp> {
+public:
+  using OpConversionPattern::OpConversionPattern;
+  LogicalResult
+  // Simply fold the operator into the pointer to the sparse storage scheme.
+  // TODO: generalize this beyond all-dense linearized "sparse" tensors
+  matchAndRewrite(ToTensorOp op, ArrayRef<Value> operands,
+                  ConversionPatternRewriter &rewriter) const override {
+    if (auto call = operands[0].getDefiningOp<CallOp>()) {
+      Value arg = call.getOperand(0);
+      if (arg.getType().isa<LLVM::LLVMPointerType>()) {
+        rewriter.replaceOp(op, arg);
+        return success();
+      }
+    }
+    return failure();
+  }
+};
+
 } // namespace
 
 /// Populates the given patterns list with conversion rules required for
 /// the sparsification of linear algebra operations.
 void mlir::populateSparseTensorConversionPatterns(TypeConverter &typeConverter,
                                                   RewritePatternSet &patterns) {
   patterns.add<SparseReturnConverter, SparseTensorToDimSizeConverter,
                SparseTensorNewConverter, SparseTensorToPointersConverter,
-               SparseTensorToIndicesConverter, SparseTensorToValuesConverter>(
-      typeConverter, patterns.getContext());
+               SparseTensorToIndicesConverter, SparseTensorToValuesConverter,
+               SparseTensorToTensorConverter>(typeConverter,
+                                              patterns.getContext());
 }

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

public:
 
   /// Bit translation.
   unsigned tensor(unsigned b) const { return b % numTensors; }
   unsigned index(unsigned b) const { return b / numTensors; }
 
   /// Returns true if bit corresponds to queried dim.
   bool isDim(unsigned b, Dim d) const { return isDim(tensor(b), index(b), d); }
 
+  /// Returns true if bit corresponds to index of output tensor.
+  bool isOutTensor(unsigned b, unsigned i) const {
+    return tensor(b) == outTensor && index(b) == i;
+  }
+
   /// Returns true if tensor access at given index has queried dim.
   bool isDim(unsigned t, unsigned i, Dim d) const {
     assert(t < numTensors && i < numLoops);
     return dims[t][i] == d;
   }
 
   /// Returns true if any set bit corresponds to queried dim.
   bool hasAnyDimOf(const llvm::BitVector &bits, Dim d) const {

 static bool findSparseAnnotations(Merger &merger, linalg::GenericOp op) {
   bool annotated = false;
   OpOperand *lhs = op.getOutputOperand(0);
   for (OpOperand *t : op.getInputAndOutputOperands()) {
     auto map = op.getTiedIndexingMap(t);
     if (!map.isProjectedPermutation())
       return false;
     auto enc = getSparseTensorEncoding(t->get().getType());
-    if (enc) {
+    if (enc)
       annotated = true;
-      if (t == lhs)
-        return false; // TODO: handle sparse outputs
-    }
     assert(map.getNumResults() == op.getRank(t));
     for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
       unsigned idx = map.getDimPosition(perm(enc, d));
+      Dim dim = toDim(enc, d);
       merger.setDim(t->getOperandNumber(), idx, toDim(enc, d));
+      // Accept only all-dense annotated "sparse" output.
+      // TODO: support truly sparse outputs too
+      if (t == lhs && dim != Dim::kDense)
+        return false;
     }
   }
   return annotated;
 }
 
 /// A DFS helper to compute a topological sort. Note that recursion is
 /// bounded by the number of implicit loops, which is always small.
 /// Returns false when a cycle is detected.

 /// dimensions. Even for dense storage formats, however, the natural index
 /// order yields innermost unit-stride access with better spatial locality.
 static bool computeIterationGraph(Merger &merger, linalg::GenericOp op,
                                   std::vector<unsigned> &topSort,
                                   bool sparseOnly) {
   // Set up an n x n from/to adjacency matrix of the iteration graph
   // for the implicit loop indices i_0 .. i_n-1.
   unsigned n = op.getNumLoops();
+  std::vector<unsigned> indegree(n, 0);
   std::vector<std::vector<bool>> adjM(n, std::vector<bool>(n, false));
 
   // Iterate over the indexing maps of every tensor in the tensor expression.
   for (OpOperand *t : op.getInputAndOutputOperands()) {
     auto map = op.getTiedIndexingMap(t);
     auto enc = getSparseTensorEncoding(t->get().getType());
     assert(map.getNumDims() == n);
     // Skip dense tensor constraints when sparse only is requested.
     if (sparseOnly && !enc)
       continue;
     // Each tensor expression and optional dimension ordering (row-major
     // by default) puts an ordering constraint on the loop indices. For
     // example, the tensor expresion A_ijk forces the ordering i < j < k
     // on the loop indices if no explicit dimension ordering is given.
     for (unsigned d = 1, rank = map.getNumResults(); d < rank; d++) {
       unsigned f = map.getDimPosition(perm(enc, d - 1));
       unsigned t = map.getDimPosition(perm(enc, d));
+      if (!adjM[f][t]) {
+        indegree[t]++;
         adjM[f][t] = true;
       }
     }
+  }
 
   // Topologically sort the iteration graph to determine loop order.
   // Report failure for a cyclic iteration graph.
   topSort.clear();
   topSort.reserve(n);
   std::vector<unsigned> visit(n, 0);
   for (unsigned i = 0; i < n; i++)
     if (visit[i] == 0)
static Value genOutputBuffer(CodeGen &codegen, PatternRewriter &rewriter,
   Value alloc = rewriter.create<memref::AllocOp>(loc, denseTp, args);
   rewriter.create<linalg::CopyOp>(loc, init, alloc);
   return alloc;
 }
 
 /// Local bufferization of all dense and sparse data structures.
 /// This code enables testing the first prototype sparse compiler.
 // TODO: replace this with a proliferated bufferization strategy
-static void genBuffers(Merger &merger, CodeGen &codegen,
+static bool genBuffers(Merger &merger, CodeGen &codegen,
                        PatternRewriter &rewriter, linalg::GenericOp op) {
   Location loc = op.getLoc();
   assert(op.getNumInputsAndOutputs() == op.getNumInputs() + 1);
   // For every tensor, find lower and upper bound on dimensions, set the
   // same bounds on loop indices, and obtain dense or sparse buffer(s).
   SmallVector<Value, 4> args;
   for (OpOperand *t : op.getInputAndOutputOperands()) {
+    unsigned tensor = t->getOperandNumber();
     auto shape = op.getShape(t);
     auto map = op.getTiedIndexingMap(t);
     auto enc = getSparseTensorEncoding(t->get().getType());
     // Scan all dimensions of current tensor.
     args.clear();
     for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
       unsigned idx = map.getDimPosition(perm(enc, d));
       // Handle sparse storage schemes.
-      if (merger.isDim(t->getOperandNumber(), idx, Dim::kSparse)) {
+      if (merger.isDim(tensor, idx, Dim::kSparse)) {
         auto dynShape = {ShapedType::kDynamicSize};
         auto ptrTp = MemRefType::get(
             dynShape, genIntType(rewriter, enc.getPointerBitWidth()));
         auto indTp = MemRefType::get(
             dynShape, genIntType(rewriter, enc.getIndexBitWidth()));
         Value dim = rewriter.create<ConstantIndexOp>(loc, d);
         // Generate sparse primitives to obtains pointer and indices.
-        codegen.pointers[t->getOperandNumber()][idx] =
+        codegen.pointers[tensor][idx] =
             rewriter.create<ToPointersOp>(loc, ptrTp, t->get(), dim);
-        codegen.indices[t->getOperandNumber()][idx] =
+        codegen.indices[tensor][idx] =
             rewriter.create<ToIndicesOp>(loc, indTp, t->get(), dim);
       }
       // Find lower and upper bound in current dimension.
       Value up;
       if (shape[d] == MemRefType::kDynamicSize) {
         up = rewriter.create<memref::DimOp>(loc, t->get(), d);
         args.push_back(up);
       } else {
         up = rewriter.create<ConstantIndexOp>(loc, shape[d]);
       }
-      codegen.sizes[idx] = codegen.highs[t->getOperandNumber()][idx] = up;
+      codegen.sizes[idx] = codegen.highs[tensor][idx] = up;
     }
-    // Perform the required bufferization. All dense inputs materialize
-    // from the input tensor. The dense output tensor needs special
-    // handling. Sparse inputs use a sparse primitive to obtain the values.
+    // Perform the required bufferization. Dense inputs materialize
+    // from the input tensors. Dense outputs need special handling.
+    // Sparse inputs use sparse primitives to obtain the values.
+    // We also accept in-place all-dense annotated "sparse" outputs.
     Type elementType = getElementTypeOrSelf(t->get().getType());
     if (!enc) {
+      // Non-annotated dense tensors.
       auto denseTp = MemRefType::get(shape, elementType);
-      if (t->getOperandNumber() < op.getNumInputs())
-        codegen.buffers[t->getOperandNumber()] =
+      if (tensor < op.getNumInputs())
+        codegen.buffers[tensor] =
             rewriter.create<memref::BufferCastOp>(loc, denseTp, t->get());
       else
-        codegen.buffers[t->getOperandNumber()] =
+        codegen.buffers[tensor] =
             genOutputBuffer(codegen, rewriter, op, denseTp, args);
     } else {
+      // Annotated sparse tensors.
+      if (tensor == op.getNumInputs() && !getInPlace(t->get()))
+        return false; // reject output if not in-place
       auto dynShape = {ShapedType::kDynamicSize};
       auto sparseTp = MemRefType::get(dynShape, elementType);
-      codegen.buffers[t->getOperandNumber()] =
+      codegen.buffers[tensor] =
           rewriter.create<ToValuesOp>(loc, sparseTp, t->get());
     }
   }
+  return true;
 }
 
 /// Constructs vector type.
 static VectorType vectorType(CodeGen &codegen, Type etp) {
   return VectorType::get(codegen.curVecLength, etp);
 }
 
 /// Constructs vector type from pointer.
static Value genTensorLoad(Merger &merger, CodeGen &codegen,
   Value val = merger.exp(exp).val;
   if (val) {
     if (codegen.curVecLength > 1 && !val.getType().isa<VectorType>())
       return genVectorInvariantValue(codegen, rewriter, val);
     return val;
   }
   // Actual load.
   SmallVector<Value, 4> args;
-  OpOperand *tensor = merger.exp(exp).e0 < op.getNumInputs()
-                          ? op.getInputOperand(merger.exp(exp).e0)
-                          : op.getOutputOperand(0);
-  auto map = op.getTiedIndexingMap(tensor);
-  auto enc = getSparseTensorEncoding(tensor->get().getType());
-  for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
-    unsigned idx = map.getDimPosition(perm(enc, d));
-    args.push_back(codegen.loops[idx]); // universal dense index
-    if (enc) {
-      args.clear();
-      args.push_back(
-          codegen.pidxs[tensor->getOperandNumber()][idx]); // position index
+  OpOperand *t = merger.exp(exp).e0 < op.getNumInputs()
+  unsigned tensor = t->getOperandNumber();
+  auto map = op.getTiedIndexingMap(t);
+  auto enc = getSparseTensorEncoding(t->get().getType());
+  unsigned rank = map.getNumResults();
+    unsigned idx = map.getDimPosition(perm(enc, rank - 1));
+    assert(codegen.pidxs[tensor][idx] != nullptr);
+    args.push_back(codegen.pidxs[tensor][idx]); // position index
+  } else {
+    for (unsigned d = 0; d < rank; d++) {
+      unsigned idx = map.getDimPosition(d);
+      args.push_back(codegen.loops[idx]); // universal dense index
     }
   }
   Location loc = op.getLoc();
-  Value ptr = codegen.buffers[tensor->getOperandNumber()];
+  Value ptr = codegen.buffers[tensor];
   if (codegen.curVecLength > 1)
     return genVectorLoad(codegen, rewriter, ptr, args);
   return rewriter.create<memref::LoadOp>(loc, ptr, args);
 }
 
 /// Generates a store on a dense tensor.
 static void genTensorStore(Merger &merger, CodeGen &codegen,
                            PatternRewriter &rewriter, linalg::GenericOp op,
-                           OpOperand *tensor, Value rhs) {
+                           OpOperand *t, Value rhs) {
   Location loc = op.getLoc();
   // Test if this is a scalarized reduction.
   OpOperand *lhs = op.getOutputOperand(0);
-  if (lhs == tensor && codegen.redVal) {
+  if (lhs == t && codegen.redVal) {
     if (codegen.curVecLength > 1)
       rhs = rewriter.create<SelectOp>(loc, codegen.curVecMask, rhs,
                                       codegen.redVal);
     codegen.redVal = rhs;
     return;
   }
   // Actual store.
   SmallVector<Value, 4> args;
-  auto map = op.getTiedIndexingMap(tensor);
-  assert(!getSparseTensorEncoding(tensor->get().getType()));
-  for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
+  unsigned tensor = t->getOperandNumber();
+  auto map = op.getTiedIndexingMap(t);
+  auto enc = getSparseTensorEncoding(t->get().getType());
+  unsigned rank = map.getNumResults();
+  if (enc) {
+    unsigned idx = map.getDimPosition(perm(enc, rank - 1));
+    assert(codegen.pidxs[tensor][idx] != nullptr);
+    args.push_back(codegen.pidxs[tensor][idx]); // position index
+  } else {
+    for (unsigned d = 0; d < rank; d++) {
       unsigned idx = map.getDimPosition(d);
       args.push_back(codegen.loops[idx]); // universal dense index
     }
-  Value ptr = codegen.buffers[tensor->getOperandNumber()];
+  }
+  Value ptr = codegen.buffers[tensor];
   if (codegen.curVecLength > 1)
     genVectorStore(codegen, rewriter, rhs, ptr, args);
   else
     rewriter.create<memref::StoreOp>(loc, rhs, ptr, args);
 }
 
 /// Generates a pointer/index load from the sparse storage scheme. Narrower
 /// data types need to be zero extended before casting the value into the
 /// index type used for looping and indexing.
 static Value genLoad(CodeGen &codegen, PatternRewriter &rewriter, Location loc,
                      Value ptr, Value s) {
   // See https://llvm.org/docs/GetElementPtr.html for some background on
   // the complications described below.
   if (codegen.curVecLength > 1) {
     // Since the index vector is used in a subsequent gather/scatter operations,
     // which effectively defines an unsigned pointer + signed index, we must
     // zero extend the vector to an index width. For 8-bit and 16-bit values,
     // an 32-bit index width suffices. For 32-bit values, zero extending the
     // elements into 64-bit loses some performance since the 32-bit indexed
     // gather/scatter is more efficient than the 64-bit index variant (if the
     // negative 32-bit index space is unused, the enableSIMDIndex32 flag can
-    // preserve this performance)). For 64-bit values, there is no good way
+    // preserve this performance). For 64-bit values, there is no good way
     // to state that the indices are unsigned, with creates the potential of
     // incorrect address calculations in the unlikely case we need such
     // extremely large offsets.
     Type etp = ptr.getType().cast<MemRefType>().getElementType();
     Value vload = genVectorLoad(codegen, rewriter, ptr, {s});
     if (!etp.isa<IndexType>()) {
       if (etp.getIntOrFloatBitWidth() < 32)
         vload = rewriter.create<ZeroExtendIOp>(
static void genLocals(Merger &merger, CodeGen &codegen,
   }
 
   // Merge dense universal index over minimum.
   if (min) {
     assert(!needsUniv);
     codegen.loops[idx] = min;
   }
 
-  // Initialize dense positions.
+  // Initialize dense positions. Note that we generate dense indices of the
+  // output tensor unconditionally, since they may not appear in the lattice,
+  // but may be needed for linearized codegen.
   for (unsigned b = 0, be = locals.size(); b < be; b++) {
-    if (locals[b] && merger.isDim(b, Dim::kDense)) {
+    if ((locals[b] || merger.isOutTensor(b, idx)) &&
+        merger.isDim(b, Dim::kDense)) {
       unsigned tensor = merger.tensor(b);
       assert(idx == merger.index(b));
       unsigned pat = at;
       for (; pat != 0; pat--)
         if (codegen.pidxs[tensor][topSort[pat - 1]])
           break;
       Value p = (pat == 0) ? rewriter.create<ConstantIndexOp>(loc, 0)
                            : codegen.pidxs[tensor][topSort[pat - 1]];
static void genStmt(Merger &merger, CodeGen &codegen, PatternRewriter &rewriter,
 
   // Wrap-up loop sequence.
   codegen.curVecLength = 1;
   genReductionEnd(merger, codegen, rewriter, op);
   genInvariants(merger, codegen, rewriter, op, exp, ldx, /*hoist=*/false);
   codegen.loops[idx] = Value();
 }
 
+/// Converts the result computed by the sparse kernel into the required form.
+static void genResult(CodeGen &codegen, PatternRewriter &rewriter,
+                      linalg::GenericOp op) {
+  RankedTensorType resType = op.getOutputTensorTypes()[0];
+  Value result = codegen.buffers.back();
+  if (getSparseTensorEncoding(resType))
+    result = rewriter.create<ToTensorOp>(op.getLoc(), resType, result);
+  else
+    result =
+        rewriter.create<memref::TensorLoadOp>(op.getLoc(), resType, result);
+  rewriter.replaceOp(op, result);
+}
+
 namespace {
 
 /// Sparse rewriting rule for generic Lingalg operation.
 struct GenericOpSparsifier : public OpRewritePattern<linalg::GenericOp> {
 public:
   GenericOpSparsifier(MLIRContext *context, SparsificationOptions o)
       : OpRewritePattern<linalg::GenericOp>(context), options(o) {}
 
LogicalResult matchAndRewrite(linalg::GenericOp op,
     // expression for the Linalg operation in SSA form.
     Operation *yield = op.region().front().getTerminator();
     Optional<unsigned> exp = buildTensorExp(merger, op, yield->getOperand(0));
     if (!exp.hasValue())
       return failure(); // build failure
 
     // Recursively generates code.
     CodeGen codegen(options, numTensors, numLoops);
-    genBuffers(merger, codegen, rewriter, op);
+    if (!genBuffers(merger, codegen, rewriter, op))
+      return failure(); // could not bufferize
     genStmt(merger, codegen, rewriter, op, topSort, exp.getValue(), 0);
-    Value result = rewriter.create<memref::TensorLoadOp>(
-        op.getLoc(), codegen.buffers.back());
-    rewriter.replaceOp(op, result);
+    genResult(codegen, rewriter, op);
     return success();
   }
 
 private:
   /// Options to control sparse code generation.
   SparsificationOptions options;
 };
 
 } // namespace
 
 /// Populates the given patterns list with rewriting rules required for
 /// the sparsification of linear algebra operations.
 void mlir::populateSparsificationPatterns(
     RewritePatternSet &patterns, const SparsificationOptions &options) {
   patterns.add<GenericOpSparsifier>(patterns.getContext(), options);
 }

mlir/test/Dialect/SparseTensor/conversion.mlir

 // RUN: mlir-opt %s --sparse-tensor-conversion | FileCheck %s
 
+#DenseVector = #sparse_tensor.encoding<{
+  dimLevelType = ["dense"]
+}>
+
 #SparseVector = #sparse_tensor.encoding<{
   dimLevelType = ["compressed"]
 }>
 
 #SparseVector64 = #sparse_tensor.encoding<{
   dimLevelType = ["compressed"],
   pointerBitWidth = 64,
   indexBitWidth = 64

 // CHECK-LABEL: func @sparse_valuesi8(
 //  CHECK-SAME: %[[A:.*]]: !llvm.ptr<i8>)
 //       CHECK: %[[T:.*]] = call @sparseValuesI8(%[[A]]) : (!llvm.ptr<i8>) -> memref<?xi8>
 //       CHECK: return %[[T]] : memref<?xi8>
 func @sparse_valuesi8(%arg0: tensor<128xi8, #SparseVector>) -> memref<?xi8> {
   %0 = sparse_tensor.values %arg0: tensor<128xi8, #SparseVector> to memref<?xi8>
   return %0 : memref<?xi8>
 }
+
+// CHECK-LABEL: func @sparse_reconstruct(
+//  CHECK-SAME: %[[A:.*]]: !llvm.ptr<i8>
+//       CHECK: return %[[A]] : !llvm.ptr<i8>
+func @sparse_reconstruct(%arg0: tensor<128xf32, #DenseVector> {linalg.inplaceable = true}) -> tensor<128xf32, #DenseVector> {
+  %0 = sparse_tensor.values %arg0 : tensor<128xf32, #DenseVector> to memref<?xf32>
+  %1 = sparse_tensor.tensor %0 : memref<?xf32> to tensor<128xf32, #DenseVector>
+  return %1 : tensor<128xf32, #DenseVector>
+}

mlir/test/Dialect/SparseTensor/dense.mlir

+// NOTE: Assertions have been autogenerated by utils/generate-test-checks.py
+// RUN: mlir-opt %s -sparsification | FileCheck %s
+
+// Test to demonstrate the difference between non-annotated dense tensors
+// and all-dense-annotated "sparse" tensors. The former class remains as
+// two-dimensional tensors that are bufferized by subsequent passes. The
+// latter class is linearized into one-dimensional buffers that are backed
+// by the runtime support library.
+
+#DenseMatrix = #sparse_tensor.encoding<{ dimLevelType = [ "dense", "dense"  ] }>
+
+#trait_2d = {
+  indexing_maps = [
+    affine_map<(i,j) -> (i,j)>,  // A
+    affine_map<(i,j) -> (i,j)>   // X (out)
+  ],
+  iterator_types = ["parallel", "parallel"],
+  doc = "X(i,j) = A(i,j) + 1"
+}
+
+#trait_3d = {
+  indexing_maps = [
+    affine_map<(i,j,k) -> (i,j,k)>,  // A
+    affine_map<(i,j,k) -> (i,j)>     // X (out)
+  ],
+  iterator_types = ["parallel", "parallel", "reduction"],
+  doc = "X(i,j) += A(i,j,k)"
+}
+
+//
+// Test with an all-dense-annotated "sparse" matrix as input and
+// a non-annotated dense matrix as output that is not inplacable.
+// This results in an explicit allocation to facilitate output.
+//
+// CHECK-LABEL:   func @dense1(
+// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>,
+// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<32x16xf32> {linalg.inplaceable = false}) -> tensor<32x16xf32> {
+// CHECK:           %[[VAL_2:.*]] = constant 1.000000e+00 : f32
+// CHECK:           %[[VAL_3:.*]] = constant 32 : index
+// CHECK:           %[[VAL_4:.*]] = constant 16 : index
+// CHECK:           %[[VAL_5:.*]] = constant 0 : index
+// CHECK:           %[[VAL_6:.*]] = constant 1 : index
+// CHECK:           %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xf32>
+// CHECK:           %[[VAL_8:.*]] = memref.buffer_cast %[[VAL_1]] : memref<32x16xf32>
+// CHECK:           %[[VAL_9:.*]] = memref.alloc() : memref<32x16xf32>
+// CHECK:           linalg.copy(%[[VAL_8]], %[[VAL_9]]) : memref<32x16xf32>, memref<32x16xf32>
+// CHECK:           scf.for %[[VAL_10:.*]] = %[[VAL_5]] to %[[VAL_3]] step %[[VAL_6]] {
+// CHECK:             scf.for %[[VAL_11:.*]] = %[[VAL_5]] to %[[VAL_4]] step %[[VAL_6]] {
+// CHECK:               %[[VAL_12:.*]] = muli %[[VAL_10]], %[[VAL_4]] : index
+// CHECK:               %[[VAL_13:.*]] = addi %[[VAL_12]], %[[VAL_11]] : index
+// CHECK:               %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_13]]] : memref<?xf32>
+// CHECK:               %[[VAL_15:.*]] = addf %[[VAL_14]], %[[VAL_2]] : f32
+// CHECK:               memref.store %[[VAL_15]], %[[VAL_9]]{{\[}}%[[VAL_10]], %[[VAL_11]]] : memref<32x16xf32>
+// CHECK:             }
+// CHECK:           }
+// CHECK:           %[[VAL_16:.*]] = memref.tensor_load %[[VAL_9]] : memref<32x16xf32>
+// CHECK:           return %[[VAL_16]] : tensor<32x16xf32>
+// CHECK:         }
+func @dense1(%arga: tensor<32x16xf32, #DenseMatrix>,
+             %argx: tensor<32x16xf32> {linalg.inplaceable = false})
+       -> tensor<32x16xf32> {
+  %c = constant 1.0 : f32
+  %0 = linalg.generic #trait_2d
+     ins(%arga: tensor<32x16xf32, #DenseMatrix>)
+    outs(%argx: tensor<32x16xf32>) {
+      ^bb(%a: f32, %x: f32):
+        %1 = addf %a, %c : f32
+        linalg.yield %1 : f32
+  } -> tensor<32x16xf32>
+  return %0 : tensor<32x16xf32>
+}
+
+//
+// Test with an all-dense-annotated "sparse" matrix as input and
+// a non-annotated dense matrix as output that is inplacable.
+// This allows updating the dense output in place.
+//
+// CHECK-LABEL:   func @dense2(
+// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>,
+// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<32x16xf32> {linalg.inplaceable = true}) -> tensor<32x16xf32> {
+// CHECK:           %[[VAL_2:.*]] = constant 1.000000e+00 : f32
+// CHECK:           %[[VAL_3:.*]] = constant 32 : index
+// CHECK:           %[[VAL_4:.*]] = constant 16 : index
+// CHECK:           %[[VAL_5:.*]] = constant 0 : index
+// CHECK:           %[[VAL_6:.*]] = constant 1 : index
+// CHECK:           %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xf32>
+// CHECK:           %[[VAL_8:.*]] = memref.buffer_cast %[[VAL_1]] : memref<32x16xf32>
+// CHECK:           scf.for %[[VAL_9:.*]] = %[[VAL_5]] to %[[VAL_3]] step %[[VAL_6]] {
+// CHECK:             scf.for %[[VAL_10:.*]] = %[[VAL_5]] to %[[VAL_4]] step %[[VAL_6]] {
+// CHECK:               %[[VAL_11:.*]] = muli %[[VAL_9]], %[[VAL_4]] : index
+// CHECK:               %[[VAL_12:.*]] = addi %[[VAL_11]], %[[VAL_10]] : index
+// CHECK:               %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_12]]] : memref<?xf32>
+// CHECK:               %[[VAL_14:.*]] = addf %[[VAL_13]], %[[VAL_2]] : f32
+// CHECK:               memref.store %[[VAL_14]], %[[VAL_8]]{{\[}}%[[VAL_9]], %[[VAL_10]]] : memref<32x16xf32>
+// CHECK:             }
+// CHECK:           }
+// CHECK:           %[[VAL_15:.*]] = memref.tensor_load %[[VAL_8]] : memref<32x16xf32>
+// CHECK:           return %[[VAL_15]] : tensor<32x16xf32>
+// CHECK:         }
+func @dense2(%arga: tensor<32x16xf32, #DenseMatrix>,
+             %argx: tensor<32x16xf32> {linalg.inplaceable = true})
+       -> tensor<32x16xf32> {
+  %c = constant 1.0 : f32
+  %0 = linalg.generic #trait_2d
+     ins(%arga: tensor<32x16xf32, #DenseMatrix>)
+    outs(%argx: tensor<32x16xf32>) {
+      ^bb(%a: f32, %x: f32):
+        %1 = addf %a, %c : f32
+        linalg.yield %1 : f32
+  } -> tensor<32x16xf32>
+  return %0 : tensor<32x16xf32>
+}
+
+//
+// Test with a non-annotated dense matrix as input and
+// an all-dense annotated "sparse" matrix as output.
+// The rewriting would fail if argx was not in-placeable.
+//
+// CHECK-LABEL:   func @dense3(
+// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<32x16xf32>,
+// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> {linalg.inplaceable = true}) -> tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> {
+// CHECK:           %[[VAL_2:.*]] = constant 1.000000e+00 : f32
+// CHECK:           %[[VAL_3:.*]] = constant 32 : index
+// CHECK:           %[[VAL_4:.*]] = constant 16 : index
+// CHECK:           %[[VAL_5:.*]] = constant 0 : index
+// CHECK:           %[[VAL_6:.*]] = constant 1 : index
+// CHECK:           %[[VAL_7:.*]] = memref.buffer_cast %[[VAL_0]] : memref<32x16xf32>
+// CHECK:           %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> to memref<?xf32>
+// CHECK:           scf.for %[[VAL_9:.*]] = %[[VAL_5]] to %[[VAL_3]] step %[[VAL_6]] {
+// CHECK:             scf.for %[[VAL_10:.*]] = %[[VAL_5]] to %[[VAL_4]] step %[[VAL_6]] {
+// CHECK:               %[[VAL_11:.*]] = muli %[[VAL_9]], %[[VAL_4]] : index
+// CHECK:               %[[VAL_12:.*]] = addi %[[VAL_11]], %[[VAL_10]] : index
+// CHECK:               %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_9]], %[[VAL_10]]] : memref<32x16xf32>
+// CHECK:               %[[VAL_14:.*]] = addf %[[VAL_13]], %[[VAL_2]] : f32
+// CHECK:               memref.store %[[VAL_14]], %[[VAL_8]]{{\[}}%[[VAL_12]]] : memref<?xf32>
+// CHECK:             }
+// CHECK:           }
+// CHECK:           %[[VAL_15:.*]] = sparse_tensor.tensor %[[VAL_8]] : memref<?xf32> to tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>
+// CHECK:           return %[[VAL_15]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>
+// CHECK:         }
+func @dense3(%arga: tensor<32x16xf32>,
+             %argx: tensor<32x16xf32, #DenseMatrix> {linalg.inplaceable = true})
+       -> tensor<32x16xf32, #DenseMatrix> {
+  %c = constant 1.0 : f32
+  %0 = linalg.generic #trait_2d
+     ins(%arga: tensor<32x16xf32>)
+    outs(%argx: tensor<32x16xf32, #DenseMatrix>) {
+      ^bb(%a: f32, %x: f32):
+        %1 = addf %a, %c : f32
+        linalg.yield %1 : f32
+  } -> tensor<32x16xf32, #DenseMatrix>
+  return %0 : tensor<32x16xf32, #DenseMatrix>
+}
+
+
+//
+// Test with a non-annotated dense matrix as input and
+// an all-dense annotated "sparse" matrix as output.
+// The rewriting would fail if argx was not in-placeable.
+// The missing innermost "k" index (due to a reduction) is accounted
+// for by scalarizing the reduction operation for the output tensor.
+//
+// CHECK-LABEL:   func @dense4(
+// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<32x16x8xf32>,
+// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> {linalg.inplaceable = true}) -> tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>> {
+// CHECK:           %[[VAL_2:.*]] = constant 8 : index
+// CHECK:           %[[VAL_3:.*]] = constant 32 : index
+// CHECK:           %[[VAL_4:.*]] = constant 16 : index
+// CHECK:           %[[VAL_5:.*]] = constant 0 : index
+// CHECK:           %[[VAL_6:.*]] = constant 1 : index
+// CHECK:           %[[VAL_7:.*]] = memref.buffer_cast %[[VAL_0]] : memref<32x16x8xf32>
+// CHECK:           %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}}>> to memref<?xf32>
+// CHECK:           scf.for %[[VAL_9:.*]] = %[[VAL_5]] to %[[VAL_3]] step %[[VAL_6]] {
+// CHECK:             scf.for %[[VAL_10:.*]] = %[[VAL_5]] to %[[VAL_4]] step %[[VAL_6]] {
+// CHECK:               %[[VAL_11:.*]] = muli %[[VAL_9]], %[[VAL_4]] : index
+// CHECK:               %[[VAL_12:.*]] = addi %[[VAL_11]], %[[VAL_10]] : index
+// CHECK:               %[[VAL_13:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_12]]] : memref<?xf32>
+// CHECK:               %[[VAL_14:.*]] = scf.for %[[VAL_15:.*]] = %[[VAL_5]] to %[[VAL_2]] step %[[VAL_6]] iter_args(%[[VAL_16:.*]] = %[[VAL_13]]) -> (f32) {
+// CHECK:                 %[[VAL_17:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_9]], %[[VAL_10]], %[[VAL_15]]] : memref<32x16x8xf32>
+// CHECK:                 %[[VAL_18:.*]] = addf %[[VAL_16]], %[[VAL_17]] : f32
+// CHECK:                 scf.yield %[[VAL_18]] : f32
+// CHECK:               }
+// CHECK:               memref.store %[[VAL_19:.*]], %[[VAL_8]]{{\[}}%[[VAL_12]]] : memref<?xf32>
+// CHECK:             }
+// CHECK:           }
+// CHECK:           %[[VAL_20:.*]] = sparse_tensor.tensor %[[VAL_8]] : memref<?xf32> to tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>
+// CHECK:           return %[[VAL_20]] : tensor<32x16xf32, #sparse_tensor.encoding<{{.*}}>>
+// CHECK:         }
+func @dense4(%arga: tensor<32x16x8xf32>,
+             %argx: tensor<32x16xf32, #DenseMatrix> {linalg.inplaceable = true})
+       -> tensor<32x16xf32, #DenseMatrix> {
+  %0 = linalg.generic #trait_3d
+     ins(%arga: tensor<32x16x8xf32>)
+    outs(%argx: tensor<32x16xf32, #DenseMatrix>) {
+      ^bb(%a: f32, %x: f32):
+        %1 = addf %x, %a : f32
+        linalg.yield %1 : f32
+  } -> tensor<32x16xf32, #DenseMatrix>
+  return %0 : tensor<32x16xf32, #DenseMatrix>
+}

mlir/test/Dialect/SparseTensor/invalid.mlir

 
 #SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed"]}>
 
 func @mismatch_values_types(%arg0: tensor<?xf64, #SparseVector>) -> memref<?xf32> {
   // expected-error@+1 {{unexpected mismatch in element types}}
   %0 = sparse_tensor.values %arg0 : tensor<?xf64, #SparseVector> to memref<?xf32>
   return %0 : memref<?xf32>
 }
+
+// -----
+
+func @sparse_to_unannotated_tensor(%arg0: memref<?xf64>) -> tensor<16x32xf64> {
+  // expected-error@+1 {{expected a sparse tensor as result}}
+  %0 = sparse_tensor.tensor %arg0 : memref<?xf64> to tensor<16x32xf64>
+  return %0 : tensor<16x32xf64>
+}
+
+// -----
+
+#SparseMatrix = #sparse_tensor.encoding<{dimLevelType = ["dense","compressed"]}>
+
+func @sparse_to_sparse_tensor(%arg0: memref<?xf64>) -> tensor<16x32xf64, #SparseMatrix> {
+  // expected-error@+1 {{unexpected non-dense dimension}}
+  %0 = sparse_tensor.tensor %arg0 : memref<?xf64> to tensor<16x32xf64, #SparseMatrix>
+  return %0 : tensor<16x32xf64, #SparseMatrix>
+}
+
+// -----
+
+#DenseMatrix = #sparse_tensor.encoding<{dimLevelType = ["dense","dense"]}>
+
+func @sparse_to_tensor(%arg0: memref<?xindex>,
+                       %arg1: memref<?xindex>,
+           %arg2: memref<?xf64>) -> tensor<16x32xf64, #DenseMatrix> {
+  // expected-error@+1 {{expected single values array}}
+  %0 = sparse_tensor.tensor %arg0, %arg1, %arg2
+    : memref<?xindex>, memref<?xindex>, memref<?xf64> to tensor<16x32xf64, #DenseMatrix>
+  return %0 : tensor<16x32xf64, #DenseMatrix>
+}

mlir/test/Dialect/SparseTensor/roundtrip.mlir

 // CHECK-LABEL: func @sparse_values(
 //  CHECK-SAME: %[[A:.*]]: tensor<128xf64, #{{.*}}>)
 //       CHECK: %[[T:.*]] = sparse_tensor.values %[[A]] : tensor<128xf64, #{{.*}}> to memref<?xf64>
 //       CHECK: return %[[T]] : memref<?xf64>
 func @sparse_values(%arg0: tensor<128xf64, #SparseVector>) -> memref<?xf64> {
   %0 = sparse_tensor.values %arg0 : tensor<128xf64, #SparseVector> to memref<?xf64>
   return %0 : memref<?xf64>
 }
+
+// -----
+
+#DenseMatrix = #sparse_tensor.encoding<{dimLevelType = ["dense","dense"]}>
+
+// CHECK-LABEL: func @sparse_to_tensor(
+//  CHECK-SAME: %[[A:.*]]: memref<?xf64>)
+//       CHECK: %[[T:.*]] = sparse_tensor.tensor %[[A]] : memref<?xf64> to tensor<16x32xf64, #{{.*}}>
+//       CHECK: return %[[T]] : tensor<16x32xf64, #{{.*}}>
+func @sparse_to_tensor(%arg0: memref<?xf64>) -> tensor<16x32xf64, #DenseMatrix> {
+  %0 = sparse_tensor.tensor %arg0 : memref<?xf64> to tensor<16x32xf64, #DenseMatrix>
+  return %0 : tensor<16x32xf64, #DenseMatrix>
+}